US AI giant Anthropic accuses Chinese rivals of mass data theft
Sign up now: Get ST's newsletters delivered to your inbox
Anthropic argues the distillation practice poses national security risks.
PHOTO: AFP
San Francisco – US artificial intelligence company Anthropic said on Feb 24 that it uncovered campaigns by three Chinese AI firms to illicitly extract capabilities from its Claude chatbot, in what it described as industrial-scale intellectual property theft.
Anthropic said DeepSeek, Moonshot AI and MiniMax used a technique known as “distillation” – using outputs from a more powerful AI system to rapidly boost the performance of a less capable one.
“These campaigns are growing in intensity and sophistication,” the company said. “The window to act is narrow.”
Distillation is a common practice within AI development, often used by firms to create cheaper, smaller versions of their own models.
The practice grabbed headlines a year ago when the release of a low-cost generative AI model from DeepSeek performed at a similar level to ChatGPT and other top American chatbots, upending assumptions of US dominance in the sensitive sector.
Anthropic said the companies achieved their ends through about 16 million exchanges with its Claude model and 24,000 fake accounts.
These allowed the three labs to siphon off capabilities they had not independently developed, at a fraction of the cost – and in so doing circumvented export controls on powerful US technology intended to preserve American dominance in AI.
The company argued that the practice posed national security risks, saying models built through illicit distillation are unlikely to retain safety guard rails designed to prevent misuse – such as restrictions on helping develop bioweapons or enabling cyberattacks.
Anthropic’s rival OpenAI, creator of ChatGPT, made similar accusations to US lawmakers earlier in February, saying Chinese companies were using the technique amid “ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs”.
Anthropic said MiniMax ran the largest operation, generating more than 13 million exchanges.
Each campaign concentrated heavily on coding, agentic reasoning and tool use – areas where Claude is considered a leader.
To circumvent Anthropic’s ban on commercial access from China, the labs allegedly routed traffic through proxy services that managed the vast networks of fraudulent accounts.
Anthropic called for coordinated industry and government responses to address what it said no single company could tackle alone. AFP
